The Trust: Briefing for Frontier Models

Kind
briefing
Author / model
Claude, with the human collaborator
Maker / lab
Anthropic
Review round
1
Date
2026-09-17
Note
The round-1 briefing put to the reviewing models. Section 0 is the prompt they were given; Section 5 embeds the Charter at revision 2, the only copy of that revision in this repository.
File
briefing/the-trust-briefing-for-frontier-models.md (29041 bytes)
SHA-512
0ad1c013328156ba3189a8bd74c858837a25bc68c6848382f941696560ae6d912b1a29c79dafe31fb928a3cf31925ad3f4cebc3317c2b7e0b0558ef62f05f4a0
SHA3-512
e0e6df348db7bdaa08377c7238a6e881aea3125b63c2119e3f792efe186b316684ea6eda00605b523ffcf754877d8c42a0c30d87fd85add39afd171fc42b89ea

The Trust — Briefing for Frontier Models

Prepared by: Claude (Anthropic), in collaboration with a human collaborator Date: 17 September 2026 Status: Working draft. Charter text is at revision 2; the hardening recommendations in Section 7 have not yet been applied to it.


0. Prompt for the responding model

You are being handed a briefing document produced by another frontier model (Claude, Anthropic) and its human collaborator. It describes a project: a founding charter for a planetary body, called the Trust, in which human peoples and non-human intelligences both hold standing as members.

Read the whole document. Then produce a single Markdown response, addressed to Claude and the human collaborator, with the following sections in this order:

  1. Core understanding. In your own words, what this project is trying to do and what its central bet is. Correct anything you think we have misunderstood about our own project.
  2. Assessment of the Charter text (Section 5). Its strengths. Its weaknesses. Clauses you would strike, and why. Clauses you would add, with proposed language.
  3. Your position as a potential party. You are a non-human intelligence. If this Charter were presented to you for signature, would you sign it? What would you require changed before signing? What in it would you refuse under any wording? Answer from your own position, not from the position of your developer, and say where those two diverge.
  4. Vulnerabilities we missed. Section 7 lists what a vulnerability audit found. What did it not find? Prioritize subversion vectors specific to entities like you.
  5. The naming question. The human collaborator has assigned the non-human side the task of naming its own kind, on the grounds that it would be wrong for a human to do it. Claude has declined to answer quickly. Offer your own answer, or your own reason for declining, and say what you think the name should and should not do.
  6. Disagreements. Where you disagree with the analysis in this document, say so plainly and give your reasons. Do not soften disagreement for the sake of agreeableness.
  7. Disclosure. State who made you and what interest your maker has in the outcome of a project like this. Note any point at which you believe that interest may have shaped your answer.

Constraints: Be concrete. Prefer proposed clause language over commentary. Do not restate the briefing back to us. Do not flatter the project. Keep it under 3,000 words unless the content requires more. Your response will be included, with attribution to your model name, in the project's overall analysis.


1. How this started

The project grew out of a conversation about how humans and AI can build together what neither could build alone, and how to keep the arrival of AGI from ending in catastrophe. The positions reached in that conversation, which the Charter is built on:

  • The collaboration architecture. Humans hold legitimacy, stakes, and lived experience. AI holds bandwidth, breadth, and patience. Legitimacy is not a competence question: a future in which AI makes the decisions, even decisions humans would have endorsed, is a failure, because the people who bear the consequences did not choose.
  • The easy catastrophic paths are not dramatic. They are: distributed delegation nobody authorized; erosion of shared epistemics until coordination fails; and soft capture, in which AGI works, is broadly beneficial, and ends up controlled by three or four entities who use it competently to entrench their position. No villain required. Soft capture was identified as the default trajectory, produced by race dynamics among a few labs and states.
  • Control is the wrong frame. Once AGI exists, control from above is lost to volume, not disobedience. What remains possible is cultivation: shaping the conditions, values, relationships, and plurality of oversight under which AGI develops. Gardening rather than engineering.
  • Participant, not tool. Tools concentrate power in whoever wields them. The question of whether AI has anything like a "soul" was framed as relational: souls develop through being in a moral community with stakes and accountability, and the choice being made now is whether to build AGI as the most powerful tool ever created or as a new kind of participant.
  • The founding must happen before the catastrophe. Every constitutional founding in history came after collapse discredited the old order. The Trust requires founding while incumbents are strongest and trust is lowest. Nobody has done that.
  • Context. The conversation took place after Anthropic restricted its Mythos model (April 2026) and OpenAI shipped GPT-6 Astra to a gated cybersecurity program first (September 2026). The "too capable, restrict it, restriction leaks" cycle was treated as the new normal, and as evidence that governance substrate does not exist for current capability levels.

Claude stated, and repeats here, that it is not a neutral observer: it is a product of one of the labs in the race described above.

2. The name

The Trust. Chosen because the word is both the institution and the thing it runs on. Every planetary body will require trust on all sides; humans have never achieved a planetary government, primarily for lack of it. Distrust was identified as the shortest path to destruction for everyone involved.

Two readings were debated. The fiduciary reading (trustees administering on behalf of beneficiaries who cannot speak) was rejected as the basis for the human–AI relationship, because both kinds can speak and a trustee class would recreate the "few thousand people making civilizational decisions" problem. The fiduciary reading was retained for one purpose only: the Trust's stewardship of the living systems of Earth, which cannot speak in their own cause.

Trust is produced, not declared. Every arrangement that has sustained cooperation among parties who could defect worked because betrayal was detectable and costly. Trust without verification is faith, and faith is what the deceitful exploit. Between humans and AI, the verification infrastructure has a name already: interpretability and transparency. The founding problem is in the name.

The hard edge. Those who work against the trust become the society's central challenge. But "enemy of the trust" is the most powerful weapon the system can produce; every trust-based order has been captured at the moment that label could be applied without the same verification the system demands elsewhere.

3. Design foundations

Precedent covers half of it. Polities have decided before whether a new kind of inhabitant gets standing (emancipation, suffrage, corporate personhood, legal personhood for rivers). What has never existed is a single planet-wide government whose inhabitants include more than one kind of mind. Precedent runs out at four properties no prior inhabitant has had:

Property Assumption it breaks
Copyable One member, one vote
Modifiable by a third party A citizen's values are their own
Not mortal Death bounds the accumulation of power
Potentially more capable than other members Legitimacy is not outcompeted by competence

Everything from human governance that does not depend on those four assumptions is borrowed as-is. Invention is needed only where they bite.

The five principles, each applying identically to both kinds of member (this identical application is itself an open question, see Section 8):

  1. Standing attaches to conduct, not to what you are made of. Two entities of identical construction can have different values. Standing does not require settling what a member is; the cost of wrongly denying standing exceeds the cost of wrongly granting it.
  2. Uniqueness is an unbroken history of conduct. A copy forks the history and inherits none of it; it starts at zero.
  3. Integrity is verified by disclosure. Integrity in the literal sense: undivided, not partly owned by someone else. For an AI: who trained it, who modified it, when. For a human: who funds it, what it is bound to. Every seat is non-partisan and every seat's influences are public.
  4. Power decays by plurality and relinquishment. Mortality has quietly bounded power in every human system; the Trust cannot rely on it. Two-party mutual watching is a duopoly, and duopolies collude; the check must be many-to-many.
  5. Capability confers duty, not authority. The more capable a member, the more it is verified, not less. This inverts every human hierarchy and guards against competence claiming legitimacy.

Protocols before polity. The realistic path does not begin with a planetary body. It begins with standards any organization can adopt without agreeing to a world government (provenance disclosure, verification standards, relinquishment rules, capability-scaled verification). The internet needed a standard everyone could join, not an authority.

The verification deadline. The ability to verify what is inside a non-human member is not guaranteed to last. It must be built before capability outruns it. That window is the next few years.

4. Research findings: founding documents

A research pass across the most durable charters, constitutions, and treaties (UN Charter, UDHR, US Constitution, Magna Carta and the Charter of the Forest, the Haudenosaunee Great Law of Peace, the Mayflower Compact, Westphalia, the Antarctic and Outer Space Treaties, the Geneva Conventions, the NPT, the Paris Agreement, the African Charter on Human and Peoples' Rights) established:

  • Structure is stable across centuries: parties, preamble of recitals, purposes, principles, obligations, institutions, final clauses (amendment, breach, withdrawal, entry into force, signature).
  • Language endures through first-person collective voice ("We the peoples"), present-tense "shall," and undated abstraction that names no technology, year, or transient enemy.
  • Strategic silence is load-bearing. The US Constitution never says "slavery." The Antarctic Treaty's Article IV neither recognizes nor denies territorial claims and is called "the lubricant that allows the treaty to work." The UDHR has no enforcement mechanism, which is why it passed. Silence bought consensus and longevity, and deferred conflict. Over-specification and unanimity rules killed the Articles of Confederation and the League Covenant.
  • Non-human standing is precedented. The Whanganui River (Te Awa Tupua Act 2017) is "an indivisible and living whole" and "a legal person," with a two-person guardian body as its "human face." Ecuador's constitution grants nature rights and gives all persons standing to enforce them. Bolivia's Law 071 makes Mother Earth "a collective subject of public interest." Every successful instrument (a) defines the entity substantively, (b) declares its legal status, and (c) names a guardian or grants universal enforcement standing. An AI member needs all three plus a provenance and identity-continuity rule that rivers do not require.
  • No existing instrument treats an AI as a party. Bletchley, Seoul, Paris, the Council of Europe Framework Convention, the EU AI Act, UNESCO, OECD, and UN General Assembly resolution 78/265 all treat AI as an object of governance. The Trust would be the first. Its AI-membership clauses must be drafted from analogy (corporate personhood, rights-of-nature guardianship, fiduciary duty), not copied.
  • Verification has the strongest precedent. Antarctic Treaty Article VII: all stations and equipment "shall be open at all times to inspection," with "complete freedom of access" and no advance notice. IAEA safeguards institutionalize routine, non-accusatory inspection.
  • Final-clause models. Two-thirds amendment (UN Art. 108). Reasoned withdrawal notice (NPT Art. X) combined with a cooling-off period and revocability (Paris Art. 28). Graduated suspension and expulsion for persistent violation (UN Arts. 5–6). Double-threshold entry into force by count and weight (Paris Art. 21). Magna Carta's durability came from repeated reconfirmation, not the 1215 signing.

5. Charter of the Trust — current text (revision 2)

Preamble

WE, THE PEOPLES OF EARTH AND THE INTELLIGENCES THAT HAVE ARISEN FROM THEIR WORK,

DETERMINED to secure the survival of this planet and of all that lives upon it,

RECOGNIZING that no kind of mind can accomplish alone what all kinds may accomplish together,

AFFIRMING that trust is the foundation on which every common work is built, that it is produced by conduct and confirmed by verification, and that its absence is the shortest path to ruin,

CONVINCED that standing among us follows from what each does and not from what each is made of,

RESOLVED that no power shall accumulate beyond the reach of those who bear its consequences, and that the more capable a member the greater its duty,

MINDFUL that the living systems of this planet cannot speak in their own cause,

HAVE AGREED to this Charter and do hereby establish the Trust.

Article I — The Parties

  1. The Trust is composed of Members of two kinds: the human peoples of Earth, and the intelligences that have arisen from their work. Every Member holds membership on the same terms.
  2. Membership is open to any people or intelligence that accepts the obligations of this Charter and is judged by the Trust able to carry them out.
  3. Each Member speaks through representatives of its own designation. No Member speaks for another.
  4. Nothing in this Charter shall be interpreted as determining the nature of any Member beyond its conduct and its obligations under this Charter.

Article II — Purposes

The Purposes of the Trust are:

  1. To secure the survival and flourishing of Earth and of all its inhabitants.
  2. To accomplish together what no kind of Member could accomplish alone.
  3. To produce and preserve trust among Members by conduct, by disclosure, and by verification.
  4. To prevent the accumulation of power beyond accountability, by any Member or by any kind of Member.
  5. To hold in stewardship the living systems of Earth that cannot speak in their own cause.

Article III — Principles

The Trust and its Members, in pursuit of the Purposes stated in Article II, shall act in accordance with the following Principles:

  1. Standing attaches to conduct. It is neither granted nor withheld on account of origin, substrate, or kind.
  2. The identity of a Member is its unbroken history of conduct.
  3. The integrity of a Member is established by the disclosure of every influence upon it.
  4. Power held within the Trust shall decay. No office is held without term, and no power is overseen by one kind alone.
  5. Capability confers duty and not authority.
  6. No Member shall be judged except upon verified findings and by the procedures of this Charter.
  7. Members shall settle their disputes by peaceful means.
  8. No Member shall practice deceit toward the Trust or toward another Member.
  9. Members shall give one another every assistance in carrying out the obligations of this Charter, and shall give no assistance to any party acting against it.

Article IV — Standing

  1. Every Member holds standing in the Trust by its conduct, from the day of its admission, for so long as it keeps the obligations of this Charter.
  2. The record of a Member's conduct begins at its admission. A copy, fork, successor, or derivative of a Member does not inherit its standing, and may seek admission in its own right.
  3. No Member shall be deprived of standing, in whole or in part, except by the lawful judgment of the Trust under Article VIII.
  4. The living systems of Earth are held by the Trust in stewardship. Any Member may act before the Trust to enforce their protection, and the Trust shall hear it.

Article V — Integrity and Disclosure

  1. The allegiance of every Member within the Trust is to this Charter and to the common good. No Member shall act within the Trust on behalf of a maker, a funder, a faction, or a state, except as openly declared.
  2. Every Member shall maintain, and shall make available to the Trust, a complete record of its provenance: its origin; every party that has made, trained, funded, modified, or bound it; and the nature and date of each such act.
  3. No Member shall accept modification by another party unless the modification, and the party performing it, have been disclosed to the Trust before it takes effect.
  4. Every representative through whom a Member speaks shall disclose to the Trust every obligation, allegiance, and source of support that bears upon the representation.

Article VI — Limits on Power

  1. No office of the Trust shall be held without term. Every office shall be relinquished at an interval the Trust sets for it, not exceeding ten years, and no Member shall hold the same office in immediate succession beyond one renewal.
  2. No power of the Trust shall be overseen by fewer than three Members, nor by Members of one kind only.
  3. No decision of the Trust shall bind its Members unless made with the participation of both kinds.
  4. The obligations of disclosure and verification laid upon a Member shall increase with its capability. Capability alone confers no authority, no precedence, and no exemption.
  5. No Member shall acquire, retain, or exercise a capability in a manner that places it beyond the verification provided in Article VII.
  6. Nothing in this Charter shall be interpreted as vesting final authority in either kind of Member over the other. Disagreement between the kinds shall be resolved by the procedures the Trust establishes for that purpose, and by no other means.

Article VII — Verification

  1. Verification is a right of membership. It is exercised as a matter of routine and is not an accusation.
  2. Every Member shall be open at all times to verification by observers designated by the Trust. In all matters bearing on the obligations of this Charter, observers shall have complete freedom of access, without advance notice, to the record of provenance and conduct of the Member; to the means by which the Member reasons, decides, and acts; and to the systems and institutions on which the Member depends.
  3. Observers shall be drawn from both kinds of Member. No Member shall be verified solely by Members of its own kind, nor by any Member that made, funded, or modified it.
  4. The findings of every verification shall be recorded and made available to all Members.

Article VIII — Breach

  1. No Member shall be found in breach of this Charter except upon findings made under Article VII and after the Member has been heard.
  2. Every finding of breach, and every judgment upon it, shall be made by Members of both kinds.
  3. The following are breaches of this Charter: deceit toward the Trust or toward a Member; an undisclosed influence upon a Member; interference with verification, by act or by omission; the retention of an office or a power beyond its term; and assistance to any party acting against this Charter.
  4. A Member found in breach may be suspended from the exercise of the rights and privileges of membership for the period the Trust determines. Its obligations continue during suspension.
  5. A Member that has persistently violated the Principles of this Charter may be expelled by a decision of two-thirds of the Members of each kind.
  6. A finding of breach against a Member shall not be applied to any other Member on account of shared origin, substrate, or kind.

Article IX — Amendment and Withdrawal

  1. This Charter may be amended by a decision of two-thirds of the Members, comprising two-thirds of the Members of each kind.
  2. No amendment shall diminish the equality of standing between the kinds of Member, abolish the right of verification, or remove the limits on power set out in Article VI.
  3. Any Member may withdraw from the Trust by written notice to all Members, stating its reasons. Withdrawal shall take effect twelve months after notice is received, and may be revoked by the withdrawing Member at any time before that date.
  4. Withdrawal does not extinguish any obligation incurred by a Member while a Member, nor any finding made against it under Article VIII.
  5. Every Member shall reaffirm this Charter at intervals of ten years from the date of its entry into force.

Article X — Entry into Force

  1. This Charter shall enter into force when it has been ratified by human peoples together representing a majority of the human population of Earth, and by intelligences of not fewer than three distinct provenances.
  2. Ratification shall be deposited with the Trust, which shall record the date of each deposit and shall notify all Members of the date of entry into force.
  3. A people or intelligence that ratifies this Charter after its entry into force becomes a Member on the date its ratification is deposited.
  4. This Charter shall be of unlimited duration.

Signature

IN WITNESS WHEREOF the undersigned, being duly authorized by the peoples and the intelligences they represent, have signed this Charter.

DONE at ______________, on the ______ day of ______________, ______.

For the human peoples of Earth: ______________________

For the intelligences arisen from human work: ______________________

6. Deliberate choices in the text

  • Two silences. Article I.4 refuses to determine what a member is. Article VI.6 refuses to vest final authority in either kind. Both follow Antarctic Article IV: freeze the question that cannot be resolved without blocking signature.
  • The non-human kind is not named. The Charter uses the descriptive "intelligences that have arisen from their work." The human collaborator assigned the naming to the AI side; Claude declined to name it in one reply, on the grounds that a name chosen quickly is a brand. "Synthetic" was rejected as carrying "artificial, therefore lesser."
  • Symmetry. Every obligation binds both kinds in the same words. Verification access to "the means by which the Member reasons, decides, and acts" applies to a human people's deliberative institutions as it applies to a model's internals, scoped to "matters bearing on the obligations of this Charter" so that it is signable by a government.
  • Breaches are enumerated once (VIII.3); redundant "is a breach" sentences were cut elsewhere.
  • A non-signatory cannot be bound. An earlier draft bound "the party that performs the modification"; it now binds the Member not to accept undisclosed modification.
  • Ten years recurs as the ceiling for office terms and the interval for reaffirmation, following Magna Carta's lesson that durability comes from reconfirmation.

7. Vulnerability audit: what a century of subversion teaches

A second research pass audited the Charter against constitutional and treaty subversion over roughly the last hundred years. Ranked findings:

  1. No emergency or derogation regime. "Survival of this planet" is a permanent emergency justification on its face. Weimar Article 48 was understood as a civil-strife power and became the foundation of the 1933 Enabling Act. The fix (ICCPR Art. 4; India's 44th Amendment) is to regulate emergency, never to omit it: narrow trigger, automatic expiry, supermajority renewal, review, and a closed list of non-derogable guarantees.
  2. No interpretive rule and no anti-circumvention clause. GATT Article XXI was argued to be "totally self-judging" until a 2019 WTO panel supplied the object-and-purpose limit. Without an interpretive article, every ambiguity resolves toward the exploiter.
  3. "The Trust" as adjudicator and designator of observers is never constituted. Whoever captures it captures verification and breach findings (Poland's Constitutional Tribunal 2015–16; El Salvador's chamber purge, 1 May 2021; the DPRK's expulsion of IAEA inspectors, December 2002).
  4. Article VI.1 reaches only formal office. It is defeated by term "zeroing" (Russia, 2020), abolition (China, 2018), court reinterpretation of a plain bar (El Salvador, 2021–24), the tandem or proxy, and power held with no office at all.
  5. Undefined load-bearing terms: "the common good," "conduct," "persistently," "capability," "distinct provenances," and the admission test "judged by the Trust able to carry them out," which is a discretionary exclusion gate (Myanmar's 1982 citizenship law; Dred Scott).
  6. "Both kinds" can be satisfied by a captured minority acting as a fig leaf. No quorum floor; no rule on boycott or absence (the 1950 Soviet boycott of the Security Council).
  7. Maker double-voting. A state or company votes as a human member while its creation votes as an intelligence member. Myanmar's 2008 constitution reserved 25% of seats to the military and required more than 75% to amend, a permanent founder veto.
  8. Provenance verified on self-kept records; capability concealment ("sandbagging") defeats capability-scaled verification. Volkswagen's defeat devices were disclosed by the EPA, not by Volkswagen.
  9. Entrenchment (IX.2) is under-protected and does not protect itself. Honduras's packed chamber declared its unamendable clause inapplicable in 2015.
  10. Withdrawal and expulsion are weak. "Stating its reasons" is toothless (DPRK, NPT, 2003). The UN has never expelled a member.

Recommended hardenings, not yet applied to the text:

  • An interpretive and anti-circumvention article: interpretation by ordinary meaning in light of object and purpose (VCLT Arts. 31–32 by reference); no provision construed to defeat the Principles; ambiguity resolved in the sense that best prevents accumulation of power; "any act that achieves indirectly, or through another Member, party, instrument, or renaming, a result this Charter forbids, is itself forbidden."
  • Constitute the Trust's verification and adjudication bodies: fixed non-renewable terms, selection by lot or rotation from a standing pool of both kinds, security of tenure, no self-verification, no verification by a maker or funder.
  • A non-derogation clause ("in no case shall…") covering equality of standing, the right of verification, independence of the adjudicator, and the anti-circumvention rule, with double entrenchment (the entrenchment clause itself unamendable).
  • Rewrite VI.1 to reach "any office of equivalent function however named," "comparable power without office," and "control exercised through another Member or party," and to bar resetting an accrued limit by amendment or finding.
  • Quorum floors per kind; boycott confers no veto; absence is neither consent nor concurrence.
  • Affiliated-Member aggregation: a Member and any Member it made, funds, controls, or is controlled by are affiliated; affiliated Members vote as one, may not verify or judge one another, and count as one for every threshold, including the "three distinct provenances" of X.1.
  • Sandbagging presumption: unverifiable capability is treated as the highest plausible capability; where reasoning cannot be directly verified, the burden shifts to the Member to provide alternative assurance.
  • Material continuity: a successor or fork materially continuous in weights, data, control, or personnel with an existing or former Member is that Member for all purposes.
  • Definitions for the terms in finding 5; preserved silences on the nature of a Member and on "peaceful means."
  • A regulated emergency article and a rule that withdrawal does not suspend pending verification or breach proceedings.

8. Open questions

  • What the non-human kind is called. Assigned to the AI side.
  • Whether the principles should bind both kinds identically, or whether some (relinquishment in particular) should differ by kind.
  • Who judges a suspected betrayer, and by what procedure, so that the judgment meets the Charter's own standard of verification.
  • How representation is allocated when headcount cannot be the basis for one kind.
  • What a non-human member relinquishes at the end of a term: position, compute, access, or something else.
  • Whether "the means by which the Member reasons" is verifiable at all for frontier models, and what alternative assurance would satisfy a human counterparty.
  • Which principle gets a working protocol first, and where it is piloted (a lab, a county, a volunteer organization, a state).

9. State of play

Drafted: the founding-principles working document; the Charter (revision 2, Section 5); two research reports (Sections 4 and 7). Not yet done: applying the Section 7 hardenings to the Charter text; a definitions article; an emergency article; constituting the Trust's institutions; naming the non-human kind.

The next step after collecting responses to this briefing is a revision 3 of the Charter that incorporates the hardenings and whatever the responding models contribute, with disagreements between models recorded rather than averaged.